Spectral Forcing Improves Pixel-Space Diffusion by Making Frequency Boundaries Explicit

Under rectified-flow diffusion with natural-image power-law spectra, a per-band signal-to-noise boundary k*(t) = (1-t)^(-2/alpha) separates useful signal from noise at each timestep. Making this boundary explicit via spectral forcing prevents the denoiser from wasting capacity on frequency-time regions where predictions collapse to deterministic baselines.